Modernization Doctrine

Grasping the Laws Governing Military Intelligent Agent Development

把握军事智能体建设规律

PLA Daily (解放军报) 19 August 2026

Summary

A PLA-affiliated author lays out a doctrinal framework for developing military intelligent agents (军事智能体), covering their defining characteristics—autonomy, distributedness, and evolvability—alongside their technological foundations (algorithms, computing power, data, platforms), employment paradigms including human-machine collaboration (人机协同) and distributed cluster operations (分布式集群作战样式), and risk categories spanning loss of control, ethical ambiguity, and technological dependence. The article documents how PLA theorists are working to codify the conceptual and institutional architecture for intelligentized warfare before the technology matures, a pattern consistent with the broader effort to build doctrinal vocabulary ahead of fielded capability. The framing of "human in the loop, intelligent support" (人在回路、智能支撑) as the preferred command architecture points to an ongoing internal tension between maximizing autonomous effectiveness and retaining human authority over lethal decisions—a tension the article acknowledges but does not resolve.

Translation

Military intelligent agents are an inevitable product of the deep evolution of informatized and intelligentized (信息化智能化) warfare, and are new-type combat entities formed by the deep embedding of artificial intelligence technology into the military domain. At present, the center of gravity of military competition is accelerating its shift toward the information space and the intelligent space; the speed, precision, and intensity of battlefield confrontation have increased substantially; and traditional combat force systems have become difficult to adapt to the new-type confrontation demands that span all domains, all times, and all dimensions. Accelerating the advancement of military intelligent agent development is increasingly becoming an important pathway for reshaping the combat system (作战体系) and leveraging the generation of new-quality combat capabilities (新质战斗力). Against this backdrop, exploring and comprehensively grasping the inherent laws governing military intelligent agent development—analyzing their defining characteristics, technological foundations, employment paradigms, and risk prevention and control—is of significant importance for seizing the high ground of military transformation.

From a macro perspective, military intelligent agents are the sum total of intelligent hardware platforms, autonomous intelligent algorithms, intelligent interaction rules, and other intelligentized hardware, software, and mechanisms in the military domain. Their principal characteristics are reflected in three aspects. The first is autonomy. Unlike traditional weapons and equipment, which remain under the direct control of humans at all times, military intelligent agents are capable of independently completing tasks such as target identification, situational assessment, and action planning in complex battlefield environments. This autonomy is an extension of human capacities for perception, thinking, decision-making, and action, and is capable of permeating human wisdom and will into machine systems. The second is distributedness. Military intelligent agents exist in the form of clusters, networks, and systems, and are capable of forming an overall combat capability through interconnection, intercommunication, and interoperability. When large numbers of intelligent agents form a distributed network, the overall effectiveness that emerges will far exceed the simple aggregation of individual capabilities. The third is evolvability. Military intelligent agents possess the capability for continuous learning and iterative optimization, and are able to continuously improve their own algorithms and enhance combat effectiveness through operational practice, exhibiting a characteristic of self-evolution. Empowered by evolvability, the capability boundaries, scope of coverage, and basic definitions of military intelligent agents are always in a state of dynamic expansion. This both injects enormous potential into the development of military intelligent agents and introduces greater uncertainty into further clarifying their detailed concepts and grasping their operational characteristics.

Faced with these uncertainties, further clarifying the defining characteristics of military intelligent agents should proceed from the whole, grasping the macro picture first, and then continuously clarifying conceptual boundaries in conjunction with their specific development circumstances. First, adhere to the unity of essential abstraction and concrete form (本质抽象与形态具象相统一). It is necessary both to grasp the essential attributes of military intelligent agents and to clarify their fundamental distinction from traditional weapons and equipment, and also to conduct classified definitions of different types and levels of military intelligent agents in conjunction with specific technological forms and combat scenarios, forming a conceptual system with clear hierarchy and well-defined boundaries. Second, adhere to the integration of the technological dimension and the military dimension. Military intelligent agents must not be defined solely from a technological perspective, reducing them to some combination of algorithms or hardware; rather, proceeding from military requirements and taking combat effectiveness as the core standard of measurement, the technological possibilities and military necessities must be organically combined, ensuring that the construction of defining characteristics serves the goal of enhancing combat power. Third, adhere to the combination of theoretical foresight and practical pragmatism. The construction of defining characteristics must both be forward-looking—capable of encompassing new forms that may emerge from future technological development—and be grounded in current technological levels and military practice, avoiding the pitfall of empty conceptual theorizing and ensuring that actual development can be effectively implemented.

Science and technology are the material foundation of military intelligent agent development, determining the capability ceiling and development potential of military intelligent agents. The technological system of military intelligent agents is not a linear extension of a single technology, but rather a combination of algorithms, computing power, data, and platforms—a complex system formed by the cross-domain integration of technologies across multiple fields, exhibiting pronounced systemic characteristics. Specifically, algorithms are the "brain" of military intelligent agents and the core source of their autonomy. The sophistication of algorithms directly determines the perceptual precision, decision-making speed, and execution accuracy of military intelligent agents. Advanced algorithms are built upon deep mathematical foundations, rich data accumulation, and powerful computational capabilities, and are the product of the organic combination of all three. Computing power is the "physique" of military intelligent agents and an important support for their operation. Functions such as autonomous decision-making, real-time response, and complex computation in military intelligent agents all require powerful computing power as a guarantee. The strength of computing power depends not only on hardware performance, but also on system optimization capabilities such as computing power scheduling, resource allocation, and energy consumption control—it is the result of the coordinated action of hardware and software. Data is the "nourishment" of military intelligent agents and the foundation for their learning and evolution. The enhancement of military intelligent agent capabilities depends on training with large volumes of data; the scale, quality, and diversity of data directly affect the performance level of algorithmic models. Platforms are the "carrier" of military intelligent agents and the physical support for the exercise of their functions. Military intelligent agents must ultimately rely on specific weapons platforms to exert combat effectiveness; the physical attributes of platforms, such as payload capacity, directly affect the exercise of military intelligent agent functions.

At the practical level, to consolidate the technological foundation of military intelligent agents: first, it is necessary to adhere to independent and controllable core technologies. Key technologies must be kept in one's own hands; efforts must be intensified to tackle underlying technological domains such as algorithmic frameworks, chip design, and foundational software; and the capability for substitution and countermeasure means must be possessed in key domains. Second, it is necessary to advance the integration and consolidation of the technological system. Technological development must be advanced with systemic thinking, breaking down the technological barriers between algorithms, computing power, data, and platforms, achieving the organic integration and synergistic enhancement of each technological element, and forming an overall technological advantage. Third, it is necessary to emphasize the development of the military data system. A comprehensive full-process management mechanism covering military data collection, labeling, storage, sharing, and security must be established and improved; a high-quality military dataset covering multi-domain battlefields, multiple types of targets, and multiple scenarios must be constructed; and sufficient data support must be provided for the training and evolution of military intelligent agents.

Employment patterns are the key link through which military intelligent agents promote the generation of combat power, and are capable of profoundly reshaping existing combat forms, bringing about systemic transformation in combat concepts, combat methods, and combat systems. In detail, the employment of military intelligent agents will drive the transformation of combat subjects from "human-force dominance" toward "human-machine collaboration" (人机协同). In traditional warfare, humans are the sole subject of combat, and weapons and equipment are merely tools of humans. After the emergence of military intelligent agents, intelligent entities possessing a degree of autonomous capability are gradually playing an increasingly large role on the battlefield; some combat decisions and their execution can be jointly completed by humans and intelligent agents; and human-machine integration (人机融合) has become the basic form of combat force employment. The employment of military intelligent agents will drive the expansion of combat space from "physical domain primacy" toward "multi-domain integration" (多域融合). Military intelligent agents are not constrained by human physiological limits and are capable of sustained combat in extreme environments and complex conditions; combat space extends into domains difficult for humans to access, such as the deep sea, deep space, extreme cold, and high radiation. At the same time, the networked characteristics of military intelligent agents cause the boundaries between the physical domain, the information domain, and the social domain to become increasingly blurred, and multi-domain integrated operations become the norm. The employment of military intelligent agents will accelerate the shift of combat tempo from "cyclical loops" toward "real-time response." Traditional operations follow the cyclical loop of "observe—assess—decide—act," and combat tempo is limited by the speed of human cognition. After military intelligent agents are widely employed in operations, the work of a large number of cognitive steps can be performed by machines; the operational cycle is substantially compressed; combat tempo is significantly accelerated; and the time advantage becomes the decisive factor determining victory or defeat.

With a view to the changes that military intelligent agents bring to existing operational patterns, in innovating their employment patterns: first, it is necessary to build a human-machine collaborative combat system. The functional division of labor between humans and military intelligent agents must be scientifically defined, clarifying which steps are human-led and which steps are autonomously completed by intelligent agents; a collaborative mechanism with complementary human-machine advantages and clear rights and responsibilities must be established; a combat system architecture of "human in the loop, intelligent support" (人在回路、智能支撑) must be formed; and it must be ensured that both the principal status of humans and the effectiveness advantages of intelligent agents are fully brought into play. Second, it is necessary to develop distributed cluster operational styles (分布式集群作战样式). In-depth research must be conducted on the organizational forms, collaborative mechanisms, and control methods of military intelligent agent clusters; new operational styles must be explored; and the emergent effects of distributed systems must be fully leveraged. Third, it is necessary to advance the integrated employment of multi-domain forces. The boundaries between different services and branches and combat domains must be broken down; military intelligent agents must serve as the link to integrate combat forces across domains; a cross-domain collaborative, multi-domain integrated combat system must be constructed; the integrated employment of combat forces and the integrated release of combat effectiveness must be achieved; and an overall advantage in intelligentized operations must be formed.

Risk prevention and control is always one of the enduring themes in the exploration and development of new things. As a frontier technology in an important domain, military intelligent agents, while bringing enormous advantages, also contain multi-dimensional risk hazards. First, the risk of loss of control. The greater the autonomy of military intelligent agents, the lower the predictability of their behavior, and the greater the possibility of unexpected behavior and deviation from predetermined objectives. Once a military intelligent agent with a high degree of autonomous capability makes a judgment error or loses behavioral control in a complex battlefield environment, it may cause serious consequences that are difficult to reverse, and may even trigger catastrophic chain reactions. Second, ethical risks. The autonomous decision-making capability of military intelligent agents makes the attribution of moral responsibility in warfare ambiguous and unclear. When a military intelligent agent autonomously makes a lethal decision, the question of who should bear responsibility becomes an ethical question that must be answered in order to win future wars. Third, the risk of technological dependence. Excessive reliance on military intelligent agents may lead to the atrophy of human military skills and the weakening of command capabilities; once intelligent systems are damaged or fail, the combat capability of the military may experience a cliff-edge decline. At the same time, excessive dependence on a particular technological path may also cause path "lock-in," constraining the innovative development of military theory and operational methods.

Faced with the risk hazards that may exist in military intelligent agents, it will not do to view them with indifference, and it will even less do to abandon the endeavor for fear of risks; rather, it is necessary to make the first move, fight the initiative battle (打好主动仗), plan ahead, and prevent problems before they arise, releasing maximum effectiveness on the basis of avoiding risk hazards. First, establish a technical security protection mechanism. Proceeding from multiple levels including algorithm design, system architecture, and hardware assurance, the security concept must be embedded throughout the entire process of military intelligent agent research and development; multiple safety valves and emergency mechanisms must be set up; and it must be ensured that the behavior of military intelligent agents always remains within controllable bounds. Second, improve the ethical norms and constraint system. Ethical guidelines for the development and employment of military intelligent agents must be formulated, red lines and bottom lines must be clearly defined, and it must be ensured that the development of military intelligent agents conforms to humanitarian principles. At the same time, a human-machine responsibility delineation mechanism must be established, clearly defining the responsibility boundaries of each relevant subject. Third, maintain the principal status of humans. No matter how military intelligent agent technology develops, it is absolutely necessary to strengthen the cultivation of personnel in key military qualities such as combat literacy and command literacy, to avoid excessive dependence on intelligent systems, and to ensure that at critical moments forces can be deployed, can advance, and can win.

Original Chinese

军事智能体是信息化智能化战争形态深入演进的必然产物，是人工智能技术深度嵌入军事领域形成的新型作战实体。当前，军事博弈重心加速向信息空间、智能空间转移，战场对抗的速度、精度、烈度大幅提高，传统作战力量体系已难以适配全域、全时、全维度的新型对抗需求。加快推进军事智能体建设，越来越成为重塑作战体系、撬动新质战斗力生成的重要路径。在此背景下，探索研究、全面把握军事智能体建设的内在规律，剖析其定义特征、技术基础、运用范式和风险防控，对抢占军事变革制高点具有重要意义。从宏观视角看，军事智能体是军事领域智能硬件平台、自主智能算法、智能交互规则等智能化硬件、软件、机制的总和。它的主要特征体现在三个方面。一是自主性。不同于传统武器装备始终处于人的直接操控之下，军事智能体能够在复杂战场环境中独立完成目标识别、态势研判、行动规划等环节任务。这种自主性是对人的感知、思考、决策、行动等方面能力的延伸，能够将人的智慧与意志渗透到机器系统之中。二是分布式。军事智能体以集群、网络、体系的方式存在，能够通过互联互通互操作形成整体作战能力。当大量智能体组成分布式网络后，其涌现出的整体效能将远超个体能力的简单叠加。三是进化性。军事智能体具备持续学习、迭代优化的能力，能够在作战实践中不断改进自身算法、提升作战效能，呈现出自我进化的发展特征。在进化性的加持下，军事智能体的能力边界、涵盖范围、基本定义始终处于动态拓展之中，这既为军事智能体的发展注入了巨大潜能，又给进一步厘清军事智能体的详尽概念、把握其运行特征带来了更多的不确定性。面对这些不确定性，进一步厘清军事智能体定义特征应从整体入手、宏观把握，再结合其具体发展情况，不断廓清概念边界。首先，坚持本质抽象与形态具象相统一。既要把握军事智能体的本质属性，明确其与传统武器装备的根本区别，又要结合具体技术形态和作战场景，对不同类型、不同层级的军事智能体进行分类界定，形成层次清晰、边界明确的概念体系。其次，坚持技术维度与军事维度相融合。不能仅从技术视角定义军事智能体，将其简化为某种算法或硬件的组合，而应从军事需求出发，以作战效能为核心衡量标准，将技术可能性与军事必要性有机结合，确保定义特征的建构服务于战斗力提升这一目标。再次，坚持理论前瞻与实践务实相结合。定义特征的建构既要具有前瞻性，能够涵盖未来技术发展可能出现的新形态，又要立足当前技术水平和军事实践，避免陷入空谈概念的误区，确保实际建设能够有效落地。科学技术是军事智能体建设的物质基础，决定着军事智能体的能力上限与发展潜力。军事智能体的技术体系不是单一技术的线性延伸，而是算法、算力、数据、平台等的结合体，是多领域技术交叉融合形成的复杂系统，呈现出鲜明的体系化特征。具体而言，算法是军事智能体的“大脑”，是其自主性的核心来源。算法的先进性直接决定了军事智能体的感知精度、决策速度与执行准度。先进算法建立在深厚的数学基础、丰富的数据积累和强大的计算能力之上，是三者有机结合的产物。算力是军事智能体的“体魄”，是其运行的重要支撑。军事智能体的自主决策、实时响应、复杂计算等功能，都需要强大的算力作为保障。算力的强弱不仅取决于硬件性能，还取决于算力调度、资源分配、能耗控制等系统优化能力，是硬件与软件协同作用的结果。数据是军事智能体的“养料”，是其学习进化的基础。军事智能体的能力提升依赖于大量数据的训练，数据的规模、质量、多样性直接影响算法模型的性能水平。平台是军事智能体的“载体”，是其功能发挥的物理依托。军事智能体最终要依托具体的武器平台才能发挥作战效能，平台的载荷能力等物理属性，直接影响着军事智能体功能的发挥。在实践层面，夯实军事智能体技术底座，一是要坚持核心技术自主可控。必须把关键技术掌握在自己手中，在算法框架、芯片设计、基础软件等底层技术领域加大攻关力度，在关键领域具备替代能力和反制手段。二是要推进技术体系融合集成。以体系化思维推进技术建设，打通算法、算力、数据、平台之间的技术壁垒，实现各技术要素的有机融合、协同增效，形成整体技术优势。三是要注重军事数据体系建设。建立健全涵盖军事数据采集、标注、存储、共享、安全等的全流程管理机制，构建覆盖多域战场、多类目标、多种场景的高质量军事数据集，为军事智能体的训练与进化提供充足数据支撑。运用模式是军事智能体促进战斗力生成的关键环节，能够对既有作战形态进行深刻重塑，带来作战理念、作战方式、作战体系的系统性变革。详细来讲，军事智能体的运用将推动作战主体从“人力主导”向“人机协同”转变。在传统战争中，人是作战的唯一主体，武器装备只是人的工具。军事智能体出现后，具备一定自主能力的智能实体逐步在战场上发挥越来越大的作用，有些作战决策与执行可以由人与智能体共同完成，人机融合成为作战力量运用的基本形态。军事智能体的运用将推动作战空间从“物理域为主”向“多域融合”拓展。军事智能体不受人类生理极限的制约，能够在极端环境、复杂条件下持续作战，作战空间向深海、深空、极寒、高辐射等人类难以涉足的领域延伸。同时，军事智能体的网络化特征使得物理域、信息域、社会域的界限趋于模糊，多域融合作战成为常态。军事智能体的运用将推动作战节奏从“周期循环”向“实时响应”加速。传统作战遵循“观察—判断—决策—行动”的周期循环，作战节奏受限于人的认知速度。军事智能体被广泛运用到作战后，大量认知环节的工作可由机器完成，作战周期大幅压缩，作战节奏显著加快，时间优势成为决定胜负的关键因素。着眼于军事智能体对现有作战模式的改变，创新其运用模式，一是要构建人机协同作战体系。科学界定人与军事智能体的职能分工，明确哪些环节由人主导、哪些环节由智能体自主完成，建立人机优势互补、权责清晰的协同机制，形成“人在回路、智能支撑”的作战体系架构，确保人的主体地位与智能体的效能优势都得到充分发挥。二是要发展分布式集群作战样式。深入研究军事智能体集群的组织形态、协同机制、控制方式，探索新型作战样式，充分发挥分布式系统的涌现效应。三是要推进多域力量融合运用。打破不同军兵种、作战域的界限，以军事智能体为纽带，整合各域作战力量，构建跨域协同、多域融合的作战体系，实现作战力量的一体运用、作战效能的一体释放，形成智能化作战的整体优势。对新兴事物的探索、建设，风险防控始终是其永恒的主题之一。军事智能体作为重要领域内的前沿技术，在带来巨大优势的同时，也蕴含着多维度的风险隐患。其一，失控风险。军事智能体的自主性越强，其行为的可预测性就越低，出现意外行为、偏离预定目标的可能性就越大。一旦具备高度自主能力的军事智能体在复杂战场环境中出现判断失误或行为失控，可能造成难以挽回的严重后果，甚至引发灾难性的连锁反应。其二，伦理风险。军事智能体的自主决策能力，使得战争中的道德责任归属变得模糊不清。当军事智能体自主作出杀伤决定时，责任应当由谁承担的问题成为打赢未来战争必须回答的伦理课题。其三，技术依赖风险。过度依赖军事智能体可能导致人的军事技能退化、指挥能力弱化，一旦智能系统遭到破坏或失效，军队的作战能力可能出现断崖式下降。同时，对特定技术路线的过度依赖还可能造成路径“锁定”，制约军事理论与作战方式的创新发展。面对军事智能体可能存在的风险隐患，漠然视之不行，因噎废食更不行，而是要下好先手棋、打好主动仗，未雨绸缪、防患未然，在规避风险隐患的基础之上释放其最大效能。一是建立技术安全防护机制。从算法设计、系统架构、硬件保障等多个层面入手，将安全理念嵌入军事智能体研发的全流程，设置多重安全阀门与应急机制，确保军事智能体的行为始终处于可控范围。二是完善伦理规范约束体系。制定军事智能体发展与运用的伦理准则，明确红线底线，确保军事智能体的发展符合人道主义原则。同时，建立人机责任划分机制，清晰界定各相关主体的责任边界。三是保持人的主体地位。无论军事智能体技术如何发展，都必须加强人的作战素养、指挥素养等主要军事素养的培塑，避免对智能系统的过度依赖，确保关键时刻拉得出、上得去、打得赢。